FSD Beta Videos (and questions for FSD Beta drivers)

[AlanSubie4Life]

FSDBeta 9.0 - 2021.4.18.12 - Unprotected Forward Facing Left Turns with Drone View

Thanks for the video! That is a very strange approach to a left turn (though I think the previous beta made a similar approach?).

You're supposed to keep the vehicle and wheels straight, so that when someone hits you from behind you go straight. Always! Seems like a major flaw in the programming.

This is definitely not the stress-free driving future for all that we have been promised, yet. Everyone in this video was very stressed. It was even stressful just watching it.

 

[mhan00]

FSDBeta 9.0 - 2021.4.18.12 - Unprotected Forward Facing Left Turns with Drone View

Tesla still needs to adjust the left turn logic so it doesn’t angle itself like that before it has a clear shot at making the turn. I was taught you’re supposed to keep the wheel straight while yielding so if you’re rear ended and your foot slips off the brake you don’t careen into traffic, plus it also doesn’t make drivers in the opposing lane of traffic nervous. The car was mostly doing okay staying out of the opposing lane, but like Chuck noted, the cars were still uncomfortable with how his car was angled and how close it was to the line.

 

[mhan00]

Thanks for the video! That is a very strange approach to a left turn (though I think the previous beta made a similar approach?).

You're supposed to keep the vehicle and wheels straight, so that when someone hits you from behind you go straight. Always! Seems like a major flaw in the programming.

This is definitely not the stress-free driving future for all that we have been promised, yet. Everyone in this video was very stressed. It was even stressful just watching it.

I will note that this version seems like a nice upgrade on these turns, even on the fails. It only stopped in the middle of the lanes once, and that looked like because a car passed in front of it. It was a weird moment, where it’s the kind of maneuver I would do as a human, seeing a small gap behind a car in the far lane so I start moving into the street to make my turn before the car is past me and slip in a little behind that car to make the gap, but it’s like the car made that decision but then still freaked out when the car passed in front of it and it stopped dead. Before on 8.2 and earlier it would constantly stop in the middle of turns and sit in the street with traffic oncoming. It’s very encouraging to see the progress, even if it isn’t as incredible as we would all hope.

 

[stopcrazypp]

Thanks for the video! That is a very strange approach to a left turn (though I think the previous beta made a similar approach?).

You're supposed to keep the vehicle and wheels straight, so that when someone hits you from behind you go straight. Always! Seems like a major flaw in the programming.

This is definitely not the stress-free driving future for all that we have been promised, yet. Everyone in this video was very stressed. It was even stressful just watching it.

Might have to do with how the car gets the best camera angles and may not be fixable. A similar move is done when making right turns.

 

[AlanSubie4Life]

Might have to do with how the car gets the best camera angles and may not be fixable. A similar move is done when making right turns.

Yeah that thought definitely crossed my mind. Not sure. Should be able to see just fine in that particular situation without doing that, though. I can see the argument for moving to the left in the reservation to clear out the median and the post from the picture and then straightening up, but the angle of the car seems to buy nothing additional.

Sometimes even as a human driver you have to break the rules a bit to get better visibility, but I do not see it as necessary in this case.

The car needs to be able to figure out how to position itself safety and optimally, somehow. Freaking out other drivers isn't cool either. Seems totally unnecessary. The wheel position is the visual cue for drivers showing that you're not going to turn into them - a observing skill that FSD will presumably add eventually to obtain human-level safety by adding it to its prediction engine (I would expect an oncoming vehicle equipped with FSD, seeing a Tesla (or any car) positioned that way, to slow down substantially and change lanes if possible, just like a good human driver).

 

[mhan00]

25:25 - this is exactly what FSD needs to do more often. If it can’t complete a maneuver safely, then it needs to move on and re-route. Preferably FSD would have the Nav calculating several routes in the background for each upcoming maneuver it can default to if it misses a turn/exit so it can seamlessly just move on to the next route. I’m surprised and gratified to see it happen here.

 

[mhan00]

https://twitter.com/x/status/1415633654819209217

FSD Beta does a smooth double lane change. I don’t think I’ve seen this before. I’m curious if this is a one off by the system (essentially a mistake of some sort) or if it is being programmed in. I’ve certainly seen a couple of scenarios in some vids where a double lane change would have been warranted to make an upcoming turn, but the car didn’t do so in the current build. Hopefully this is intended and becomes more consistent moving forward.

 

[linux-works]

It looks like their RADAR hardware uses CAN bus (1 Mbps, or *maybe* 10 Mbps). I was assuming it used something much faster, like automotive Ethernet. That design also means Tesla can't ever retrofit LIDAR or 4D RADAR over that connection. *sigh*

If those CAN buses are driven by a PLC that can be reprogrammed to treat the input as an LVDS camera feed, then it's trivial, but failing that, you'd have to do something a bit more clever for now. It is generally believed that the wide-field camera is useless. And when you're deciding whether it is safe to make a turn, the main rear camera is also basically useless. So:

• Repurpose the CAN bus from the SDC to drive a small PLC that acts like a 4x2 routing switcher, in which each of the two inputs gets video data from either the LVDS feed from one of existing cameras above or from one of the newly added cameras.
• Repurpose the wiring from the RADAR unit to the SDC so that it becomes two LVDS feeds from the new cameras.
• Add short LVDS cables from the routing switcher to the SDC.
• Add an electrical source somewhere.
• Mount the routing switcher somewhere.

And you're done except for software. Then, in the next SDC revision, add more LVDS ports and drop the switcher.

you're talking my language, now, lol.

one thing to think about: if the device that was on CAN segment X was a mid speed device and it goes up to a higher data density, you would likely need to 'rebalance' the can bus hier tree so that nothing gets starved.

I like automotive ethernet better, anyway. its the future, as CAN bus is dying (slowly, but surely; at least for the more complex ecu's).

 

[linux-works]

Yes, the driver should keep their eyes on the road at all times and not try to "drive just with the screen".

I'd actually kind of go further; there needs to be 2 people in the car if fsd beta is being tested. sorry, but I feel that way about this and so should lots of you.

one ENTHUSIAST who is trying to show tesla and his new shiny toy in the best light is not paying enough attention to ANYTHING. they are overwhelmed with inputs and not only that, they are talking at the same time. trying to advertise their channels and all that. dont like this aspect of the public beta at all. personally, no one should be ENCOURAGED to run cameras and be a youtuber. this is not the way to run a safe public beta test on public roads, dammit.

sorry to rain on y'alls parade, here; and I know I'm in the minority here. everyone wants to SEE this evolve. I want to, too; but not like this.

a responsible ceo would stop this behavior and run proper beta tests. yes, with less publicity. its how it should be done.

eventually, one of the YTers is gonna hit someone and that will be the end. lawmakers may even go quite far to put us all backwards. the red states would just LOVE to see this experiment fail.

I hope tesla rethinks how this is being done and makes serious changes. this is NOT an advertising stunt, guys. grow the hell up, tesla. stop being frat boys.

 

[dgatwood]

you're talking my language, now, lol.

one thing to think about: if the device that was on CAN segment X was a mid speed device and it goes up to a higher data density, you would likely need to 'rebalance' the can bus hier tree so that nothing gets starved.

I like automotive ethernet better, anyway. its the future, as CAN bus is dying (slowly, but surely; at least for the more complex ecu's).

To clarify, what I was proposing there (assuming it's not a PLC with dedicated wiring for talking to a single device) was taking a pretty fast CAN bus device and turning it into a single-digit-bits per minute device that literally gets a boolean flag sent once in a while to switch between "enable camera 3" mode and "enable camera 1" mode. The high-speed data would go through the existing camera input.

 

[Knightshade]

https://twitter.com/x/status/1415633654819209217

FSD Beta does a smooth double lane change. I don’t think I’ve seen this before. I’m curious if this is a one off by the system (essentially a mistake of some sort) or if it is being programmed in. I’ve certainly seen a couple of scenarios in some vids where a double lane change would have been warranted to make an upcoming turn, but the car didn’t do so in the current build. Hopefully this is intended and becomes more consistent moving forward.

https://twitter.com/x/status/1415782123298902019

FSD beta 9 is using the current regular production vision-only highway code- not any special advanced stuff.

 

[diplomat33]

I'd actually kind of go further; there needs to be 2 people in the car if fsd beta is being tested. sorry, but I feel that way about this and so should lots of you.

one ENTHUSIAST who is trying to show tesla and his new shiny toy in the best light is not paying enough attention to ANYTHING. they are overwhelmed with inputs and not only that, they are talking at the same time. trying to advertise their channels and all that. dont like this aspect of the public beta at all. personally, no one should be ENCOURAGED to run cameras and be a youtuber. this is not the way to run a safe public beta test on public roads, dammit.

sorry to rain on y'alls parade, here; and I know I'm in the minority here. everyone wants to SEE this evolve. I want to, too; but not like this.

a responsible ceo would stop this behavior and run proper beta tests. yes, with less publicity. its how it should be done.

eventually, one of the YTers is gonna hit someone and that will be the end. lawmakers may even go quite far to put us all backwards. the red states would just LOVE to see this experiment fail.

I hope tesla rethinks how this is being done and makes serious changes. this is NOT an advertising stunt, guys. grow the hell up, tesla. stop being frat boys.

Agree 1000%. There is a good reason why AV companies use professional safety drivers and 2 people in the car.

Here is yet another example that I think supports your case:

https://twitter.com/x/status/1415779767631163393

 

[mhan00]

https://twitter.com/x/status/1415782123298902019

FSD beta 9 is using the current regular production vision-only highway code- not any special advanced stuff.

But the car was clear;y on FSD Beta when it made the double lane change. You can seen the display with the FSD Beta UI. It changed to NoAP after it completed the maneuver.

 

[AndyH]

Thanks for sharing. FSD Beta definitely did a lot of things really well. There were some cases that FSD Beta did not do so well. City FSD can be very challenging.

The garbage truck scenario is actually a pretty typical problem that many AVs struggle with. It seems obvious to humans to go around but it is not obvious to AVs. Cruise also showed examples where they needed remote assistance to tell the AV what to do in a similar situation. The Cruise AV would wait behind a double parked delivery truck expecting it to move, instead of going around it. This is a good example of how city FSD can be challenging for AVs and why prediction and planning are more difficult problems than perception. City driving is full of situations where the challenge is not so much detecting objects but figuring out what lots of other road users are doing and then making a decision about what the car should do.

I wouldn't necessarily judge something to be a problem based on how development vehicles handle a situation. Developers are likely overly cautious and limit situations such as this when an AV needs to drive on the wrong side of the road.

The near term goal when developing AVs is also not to create sentinent beings that make decisions. AVs will not solve problems by using deductive reasoning. AVs are machines that follow patterns in NNs conjured up by a super computer. The challenge for Tesla is to develop NNs with patterns that safely gets a robotaxi through a high enough percentile of situations to be financially viable. It wont be perfect from the get go and robotaxis will likely get stuck a lot in the beginning. But they will improve and I also think it's likely that both traffic systems and road user behaviour will be adapted when robotaxis begin to take over.

 

[ahoen117]

Agree 1000%. There is a good reason why AV companies use professional safety drivers and 2 people in the car.

Here is yet another example that I think supports your case:

https://twitter.com/x/status/1415779767631163393

This is an odd situation.. The left signal seems to be because the nav wants you to go to the left. But because of the divided lanes you cant turn left (there's also a no left turn sign on the pillar). It's interesting though that the turn signal seems to be coming from maps/nav data, but the rest of the system was smart enough not to attempt to turn left across the other lane in the middle of that intersection.

Don't get me wrong, this defiantly shows an issue with lane choice but its interesting to try and break down the situation instead of just complaining about it.

I also see the point about multiple people. Its easy to see this info after the fact from a video, but trying to pay attention to the world around, the map data, and the predicted path is a lot to take in.

 

[diplomat33]

This is an odd situation.. The left signal seems to be because the nav wants you to go to the left. But because of the divided lanes you cant turn left (there's also a no left turn sign on the pillar). It's interesting though that the turn signal seems to be coming from maps/nav data, but the rest of the system was smart enough not to attempt to turn left across the other lane in the middle of that intersection.

Don't get me wrong, this defiantly shows an issue with lane choice but its interesting to try and break down the situation instead of just complaining about it.

I also see the point about multiple people. Its easy to see this info after the fact from a video, but trying to pay attention to the world around, the map data, and the predicted path is a lot to take in.

Oh I am fine with trying to analyze the videos to troubleshoot what is going on. I am not just complaining for the sake of complaining or being a "hater". But the point of my post is that yes, it is a lot to take in which is why you really need 2 people both attentive when testing autonomous driving. To have only 1 untrained driver who is also recording for his youtube channel, is problematic IMO, especially since FSD Beta is very much a work in progress and can fail in unexpected ways as Elon even admitted. Ultimately, I want FSD Beta to get better in a safe way.

 

[techlogik]

^This of course as other's have posted is ridiculous. I know they don't really teach people anymore how to drive and test this scenario on driver's test. Example, as another poster mentioned, you should never have your car/wheels angled towards oncoming traffic in this scenario. If you get rear-ended, you will be jettisoned into oncoming traffic head on!!! That is deadly. At least with wheels/car straight, you would likely go into the median forward...and not die. Just plain dangerous. No other way to put it.

Next, the fact that it turns the wheels/car towards oncoming traffic AND slowly creeps and makes it seem like it is about to launch into head on traffic or not even stop until the line. Again, unnerving and dangerous.

Two steps forward 1 step back in many respects based on videos/scenarios.

 

[ahoen117]

Oh I am fine with trying to analyze the videos to troubleshoot what is going on. I am not just complaining for the sake of complaining or being a "hater". But the point of my post is that yes, it is a lot to take in which is why you really need 2 people both attentive when testing autonomous driving. To have only 1 untrained driver who is also recording for his youtube channel, is problematic IMO, especially since FSD Beta is very much a work in progress and can fail in unexpected ways as Elon even admitted. Ultimately, I want FSD Beta to get better in a safe way.

Oh, I'm so sorry. The complaining comment was meant towards the author of the tweet, not you. There are some people on here who seem to complain for the fun of it but you're not one of them.

 

[Bladerskb]

This is an odd situation.. The left signal seems to be because the nav wants you to go to the left. But because of the divided lanes you cant turn left (there's also a no left turn sign on the pillar). It's interesting though that the turn signal seems to be coming from maps/nav data, but the rest of the system was smart enough not to attempt to turn left across the other lane in the middle of that intersection.

Don't get me wrong, this defiantly shows an issue with lane choice but its interesting to try and break down the situation instead of just complaining about it.

I also see the point about multiple people. Its easy to see this info after the fact from a video, but trying to pay attention to the world around, the map data, and the predicted path is a lot to take in.

Are you missing the part where its turning right into opposing traffic with an imminent head on collision?

 

[Bladerskb]

What's under-appreciated about V9 is that Tesla has essentially solved vision and is now just working on gathering more diverse data and labeling everything they can.

Why I think they've solved vision:
1) 3D environment and objects within are extremely stable, and V9 only shows what it sees with very low latency. For example, it only needs to see a small slice of a car to know it's there, its orientation, how far it is, and how fast it's moving.
2) V9 sees brake lights and associates them with the correct cars. Even when it only sees 1 brake light, it correct assumes that both the car's brake lights are on.
3) Visualization shows that V9 can see cars in all their various orientations, smoothly.
4) V9 can see and makes visual inferences (is it moving? is parked? should I be concerned? etc.) on ~40+ objects at a time

Karpathy mentioned recently that he's narrowing in on a labeling workflow that consistently produces better and better results. V9 is a demonstration of how powerful this workflow and "4D" video labeling is.

Although we see a lot of V9 "fails," I think we're very close to an update where everything seems to click, and we get surprisingly good performance.

Almost 9 months ago you were touting that Tesla will have L5 in 6-9 months with ~150k miles per safety disengagement on avg and that it was "game over" for the competition and that it would put them "10 years ahead"... Now you are saying that Tesla solved vision with <10 miles per safety disengagement on avg

The mental gymnastics of Tesla fans never ceases to amaze me.